arXiv subjects

Poppy Zhang

Publications and source records attributed to Poppy Zhang.

6 recordsLinked to original sources

arXiv2026

Offline-to-Online Creative Optimization with Generative Models and Adaptive Testing

Ad creative optimization is increasingly constrained by evaluation rather than generation. Generative models can produce many plausible creatives, but reliable evaluation requires online experiments, in which only a limited slate can be tested. We study how to use data from historical A/B tests to generate and select the candidates in that slate. We developed and deployed a performance-driven offline-to-online workflow that guides creative generation with a predictive model as an inference-time critic. In the offline phase, we use a predictive model trained on historical experiments to rank and refine variants created by a generative model. A final test slate is then deployed in an online adaptive experiment. In a 50-arm field experiment, we found that the best creative generated with this method yielded 45.1% higher engagement than the best human-authored creative. Two additional experiments showed the same upper-tail pattern, with lifts of 46.7% and 36.2%. We found that despite the predictive model being too noisy to directly identify the best creative offline, it effectively guides the generative model toward creating strong candidates that can be efficiently evaluated in an adaptive experiment. The results suggest a design principle for creative optimization with generative models: use predictive models to guide generation of a slate to test, judge the slate by whether it contains high-performing candidates at a feasible test size, and use adaptive experiments to select among candidates while limiting traffic lost to weak arms.

cs.AI↗

arXiv2026

Decoding the Hook: A Multimodal LLM Framework for Analyzing the Hooking Period of Video Ads

Video-based ads are a vital medium for brands to engage consumers, with social media platforms leveraging user data to optimize ad delivery and boost engagement. A crucial but under-explored aspect is the 'hooking period', the first three seconds that capture viewer attention and influence engagement metrics. Analyzing this brief window is challenging due to the multimodal nature of video content, which blends visual, auditory, and textual elements. Traditional methods often miss the nuanced interplay of these components, requiring advanced frameworks for thorough evaluation. This study presents a framework using transformer-based multimodal large language models (MLLMs) to analyze the hooking period of video ads. It tests two frame sampling strategies, uniform random sampling and key frame selection, to ensure balanced and representative acoustic feature extraction, capturing the full range of design elements. The hooking video is processed by state-of-the-art MLLMs to generate descriptive analyses of the ad's initial impact, which are distilled into coherent topics using BERTopic for high-level abstraction. The framework also integrates features such as audio attributes and aggregated ad targeting information, enriching the feature set for further analysis. Empirical validation on large-scale real-world data from social media platforms demonstrates the efficacy of our framework, revealing correlations between hooking period features and key performance metrics like conversion per investment. The results highlight the practical applicability and predictive power of the approach, offering valuable insights for optimizing video ad strategies. This study advances video ad analysis by providing a scalable methodology for understanding and enhancing the initial moments of video advertisements.

cs.MM↗

arXiv2026

MLLM-VADStory: Domain Knowledge-Driven Multimodal LLMs for Video Ad Storyline Insights

We propose MLLM-VADStory, a novel domain knowledge-guided multimodal large language models (MLLM) framework to systematically quantify and generate insights for video ad storyline understanding at scale. The framework is centered on the core idea that ad narratives are structured by functional intent, with each scene unit performing a distinct communicative function, delivering product and brand-oriented information within seconds. MLLM-VADStory segments ads into functional units, classifies each unit's functionality using a novel advertising-specific functional role taxonomy, and then aggregates functional sequences across ads to recover data-driven storyline structures. Applying the framework to 50k social media video ads across four industry subverticals, we find that story-based creatives improve video retention, and we recommend top-performing story arcs to guide advertisers in creative design. Our framework demonstrates the value of using domain knowledge to guide MLLMs in generating scalable insights for video ad storylines, making it a versatile tool for understanding video creatives in general.

cs.MM↗

arXiv2025

LLM Prompt Duel Optimizer: Efficient Label-Free Prompt Optimization

Large language models (LLMs) are highly sensitive to prompts, but most automatic prompt optimization (APO) methods assume access to ground-truth references (e.g., labeled validation data) that are costly to obtain. We propose the Prompt Duel Optimizer (PDO), a sample-efficient framework for label-free prompt optimization based on pairwise preference feedback from an LLM judge. PDO casts prompt selection as a dueling-bandit problem and combines (i) Double Thompson Sampling to prioritize informative comparisons under a fixed judge budget, with (ii) top-performer guided mutation to expand the candidate pool while pruning weak prompts. Experiments on BIG-bench Hard (BBH) and MS MARCO show that PDO consistently identifies stronger prompts than label-free baselines, while offering favorable quality--cost trade-offs under constrained comparison budgets.

cs.CL↗

arXiv2025

Characterizing and Minimizing Divergent Delivery in Meta Advertising Experiments

Many digital platforms offer advertisers experimentation tools like Meta's Lift and A/B tests to optimize their ad campaigns. Lift tests compare outcomes between users eligible to see ads versus users in a no-ad control group. In contrast, A/B tests compare users exposed to alternative ad configurations, absent any control group. The latter setup raises the prospect of divergent delivery: ad delivery algorithms may target different ad variants to different audience segments. This complicates causal interpretation because results may reflect both ad content effectiveness and changes to audience composition. We offer three key contributions. First, we make clear that divergent delivery is specific to A/B tests and intentional, informing advertisers about ad performance in practice. Second, we measure divergent delivery at scale, considering 3,204 Lift tests and 181,890 A/B tests. Lift tests show no meaningful audience imbalance, confirming their causal validity, while A/B tests show clear imbalance, as expected. Third, we demonstrate that campaign configuration choices can reduce divergent delivery in A/B tests, lessening algorithmic influence on results. While no configuration guarantees eliminating divergent delivery entirely, we offer evidence-based guidance for those seeking more generalizable insights about ad content in A/B tests.

econ.GN↗

arXiv2023

Node Attribute Prediction on Multilayer Networks with Weighted and Directed Edges

With the rapid development of digital platforms, users can now interact in endless ways from writing business reviews and comments to sharing information with their friends and followers. As a result, organizations have numerous digital social networks available for graph learning problems with little guidance on how to select the right graph or how to combine multiple edge types. In this paper, we first describe the types of user-to-user networks available across the Facebook (FB) and Instagram (IG) platforms. We observe minimal edge overlap between these networks, indicating users are exhibiting different behaviors and interaction patterns between platforms. We then compare predictive performance metrics across various node attribute prediction tasks for an ads click prediction task on Facebook and for a publicly available dataset from the Open Graph Benchmark. We adapt an existing node attribute prediction method for binary prediction, LINK-Naive Bayes, to account for both edge direction and weights on single-layer networks. We observe meaningful predictive performance gains when incorporating edge direction and weight. We then introduce an approach called MultiLayerLINK-NaiveBayes that can combine multiple network layers during training and observe superior performance over the single-layer results. Ultimately, whether edge direction, edge weights, and multi-layers are practically useful will depend on the particular setting. Our approach enables practitioners to quickly combine multiple layers and additional edge information such as direction or weight.

cs.SI↗